A deep learning-based framework for fundus image classification of inherited retinal degeneration in dogs

Inherited retinal degenerations (IRDs) are a frequent cause of vision impairment in numerous breeds of dogs. Veterinary ophthalmologists frequently suspect retinal degeneration when funduscopic examinations reveal increased tapetal reflectivity and retinal vascular attenuation; however, these clinical signs can be subtle, especially in early disease. In this study, we evaluated three deep learning architectures using fundus images to detect degenerative retinal disease in dogs. An EfficientNetV2-based convolutional neural network with a custom classification head was selected for final evaluation. Using dog-level analysis on an independent held-out test set, the model achieved 87.5% accuracy, 87.5% sensitivity, 87.5% specificity, and an area under the receiver operating characteristic curve of 0.932. Dog-level calibration analysis yielded a Brier score of 0.093. Under constrained single-image evaluation conditions, the model showed higher dog-level classification performance than the majority-vote classification of three experienced veterinary ophthalmologists, primarily by reducing false-positive classification of normal dogs as affected. Exact McNemar testing demonstrated a significant paired difference between the model and expert aggregate classifications at the dog level (p = 0.000122). Gradient-weighted Class Activation Mapping (Grad-CAM) was incorporated to provide qualitative visualization of retinal image regions associated with model predictions. These results support the potential of deep learning as an objective decision-support tool for canine fundus image interpretation, although broader multi-center and prospective validation is needed before clinical deployment.